DeepSomatic: Accurate somatic small variant discovery for multiple sequencing technologies

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Somatic variant detection is crucial for cancer genomics.
• Long-read sequencing offers advantages over short-read for repeat mapping and variant phasing.
• Existing methods often struggle with specific variant types or data formats.

Purpose of the Study:

• Introduce DeepSomatic, a deep learning method for somatic variant detection.
• Evaluate DeepSomatic's performance across short-read and long-read sequencing technologies.
• Address the need for public training and benchmarking data in somatic variant detection.

Main Methods:

• Developed DeepSomatic, a deep learning model for Single Nucleotide Variants (SNVs) and insertions/deletions (indels).
• Applied DeepSomatic to whole-genome and exome sequencing data from tumor-normal, tumor-only, and FFPE samples.
• Generated and released a dataset of five matched tumor-normal cell line pairs sequenced using Illumina, PacBio HiFi, and Oxford Nanopore Technologies.

Main Results:

• DeepSomatic demonstrates superior performance in detecting somatic SNVs and indels compared to existing callers.
• The method shows consistent accuracy across different sequencing technologies (short-read and long-read).
• Performance is particularly strong for indel detection.

Conclusions:

• DeepSomatic is a robust and versatile tool for somatic variant detection.
• The open-source dataset facilitates further research and development in cancer genomics.
• DeepSomatic advances the analysis of both short-read and long-read sequencing data for cancer research.